Skip to main content

Fudan's Final Exam Flips the Script: Students Stump AI, 4 Get Perfect Scores

Fudan University's 'Data Mining Technology' course just wrapped up a final exam that turned the tables on traditional testing. Instead of sitting in a classroom answering questions, students became the question setters—and their target was AI.

The New Exam Format

Each student had to independently design 10 data mining calculation problems, each with a unique correct answer and a complete derivation process. Then, they used these questions to test three AI models at different ability levels. The scoring logic was simple but clever: the more AI models the questions stumped, and the harder the questions that AI got wrong, the higher the score.

Image

Why the Change?

Professor Xiao Yanghua, the course leader from Fudan's School of Computing and Intelligence Innovation, explained that traditional exam questions have lost their edge in the AI era. If teachers still set standard algorithm problems, AI's speed and accuracy easily surpass any human student. Competing head-on with AI in its strong suits is pointless, he argued.

So he flipped the script. The core idea: guide students to deeply understand professional knowledge so they can spot AI's logical blind spots. It's not about outrunning AI—it's about outsmarting it.

The Results

Out of 51 final answer sheets submitted, the results were dramatic. 50 out of 50 students managed to stump at least one AI model. Only one student failed to trip up any AI. The class average score was 85.7 points.

The three models under fire were DeepSeek V4-Flash, MiniMax M2.7, and Claude Sonnet 4.6. The top-tier Claude model proved resilient—no student managed to completely stump it. But 4 students created high-level original questions that made the weaker AI model score zero on the entire test paper, earning them full marks.

Key Points

  • Innovative exam format: Students designed questions to test AI, not the other way around.
  • Scoring logic: The harder the AI was stumped, the higher the student's score.
  • Results: 50 out of 50 students stumped at least one AI; 4 earned full marks by making a weaker AI score zero.
  • Top AI unbeaten: Claude Sonnet 4.6 was not completely stumped by any student.
  • Teaching shift: Reflects a move away from traditional exams in the AI era.